{"data":{"kind":"file","path":"README.md","version_id":"ckyhq3k2dac91rmh3ytgejss","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":3260,"modified_at":"2026-11-07T15:31:55.502000","content_hash":"f35a3e3d3e1b4722bf20fcb43f1cda8e011fd79924a4615538e69db5ab638fb8"},"entries":[],"content":"# mini-swe-prime — Prime Intellect Environments Hub\n\nA minimal **Verifiers v1** environment grounded in the public [mini-SWE-agent v2](https://github.com/SWE-agent/mini-swe-agent) source and documentation.\n\n## What this evaluates\n\nThe model answers precise questions on the `mini` CLI, install commands, and source architecture. Exact-match scoring (after basic whitespace/case/backtick normalization) gives reproducible rewards of 0 or 1. Tasks are deterministically split into training (8) and evaluation (2) examples.\n\n**Important scope:** This baseline does **not** run mini-SWE-agent itself, execute shell commands, patch GitHub issues, score tests, or make sandbox API calls. It is a small, auditable source-understanding evaluation—not SWE-bench. Do not advertise it as an agentic coding benchmark.\n\n## Structure\n\nThe package ships one Verifiers v1 plugin module, `mini_swe_prime`, which exports:\n\n- `MiniSWETaskset` / `MiniSWETasksetConfig` — the fixed question set with a `split` knob (`train`, `eval`, `all`).\n- `MiniSWETask` / `MiniSWETaskData` — one question, its reference answer, and its category.\n- `MiniSWEHarness` / `MiniSWEHarnessConfig` — the taskset's default harness: a tool-less chat loop (`null` subclass), so the default rollout needs no container.\n- `MiniSWEEnv` / `MiniSWEEnvConfig` — the single-agent env; one turn per episode.\n- `load_taskset`, `load_harness`, `load_environment` — the loader entry points.\n\n## Installation\n\nRequires Python 3.11+, [`uv`](https://docs.astral.sh/uv/), and the [Prime CLI](https://github.com/PrimeIntellect-ai/prime).\n\n```bash\nuv tool install prime\nprime login\nuv pip install -e .\n```\n\nInstall the published environment from the Hub:\n\n```bash\nprime env install mini-swe-prime\n```\n\n## Verify and evaluate\n\n```bash\npython -m compileall -q mini_swe_prime\npython -m unittest discover -s tests -v\nvf-eval mini-swe-prime -n 2 -r 1 --dry-run --no-push\n```\n\n`vf-eval` requires a configured inference provider and model; see `vf-eval --help` for your installed version. No API key is required by the environment itself. This package has not been run against a live Prime evaluation endpoint here.\n\n## Publish\n\n```bash\nprime login\nprime env push mini-swe-prime\n```\n\nThe exact CLI arguments can change across Prime CLI releases; check `prime env push --help`. Publication requires your Prime account.\n\n## Files\n\n- `mini_swe_prime/__init__.py` — v1 taskset, harness, loaders, reward\n- `pyproject.toml` — package and eval metadata\n- `tests/test_scoring.py` — dependency-light scoring and taskset checks\n- `SOURCE.md` — upstream source and implementation notes\n\n## Future: full mini-SWE-agent coding benchmark\n\nTo run actual code-repair rollouts, implement a `CliAgentEnv` / sandbox-backed harness with an image containing a pinned mini-swe-agent commit, per-task resettable repositories, agent invocation and interception, test-based rewards, and trajectory/error reporting. That needs integration tests with Prime Sandboxes and should be published as a separate environment version.\n\n## Credits\n\nSource: https://github.com/SWE-agent/mini-swe-agent (original project by the mini-SWE-agent contributors). This environment is an independent benchmark wrapper, not an official upstream project.\n","encoding":"utf-8","truncated":false,"total_bytes":3260},"status":null}